refactor: remove unused other_mimi MimiModel instance (~200MB memory savings)

[mvanhorn]

Summary

Removes the second MimiModel instance (other_mimi) from server.py and offline.py. Every call to other_mimi.encode() and other_mimi.decode() discards the result by assigning to _. The primary mimi instance produces all actual output. Removing other_mimi frees ~200MB GPU memory.

Why this matters

#46 identified that other_mimi loads a full copy of the Mimi codec weights but never uses any output. On memory-constrained deployments (consumer GPUs, Jetson), 200MB is significant.

Changes

moshi/moshi/server.py:

• Removed other_mimi field from ServerState dataclass
• Removed other_mimi parameter from __init__, warmup, and handle_chat
• Removed other_mimi = loaders.get_mimi(...) instantiation in main()
• Removed all _ = self.other_mimi.encode(chunk) and _ = self.other_mimi.decode(tokens[:, 1:9]) calls

moshi/moshi/offline.py:

• Removed other_mimi parameter from warmup() and decode_tokens_to_pcm()
• Removed other_mimi = loaders.get_mimi(...) instantiation in run_inference()
• Removed all _ = other_mimi.encode(chunk) and _ = other_mimi.decode(tokens[:, 1:9]) calls

Net: -22 lines, +6 lines (signature adjustments).

Testing

Verified via grep -rn other_mimi --include="*.py" that zero references remain. The primary mimi encode/decode path is untouched.

No test suite exists in this repo.

Note for maintainer

If other_mimi was intended for a future streaming state feature or a separate voice prompt encoding path, please let me know and I'll revert. Based on the current code, all its outputs are discarded.

Fixes #46

This contribution was developed with AI assistance (Claude Code).